Image Cutout Algorithm Based On Local Similarity

Image segmentation is an important research topic in digital image processing. Object cutout is to extract the semantic object from the image, which not only depends on low-level information of image, but also needs high-level information of user-interface. Good algorithms of image cutout should extract the contour of the objects what users desire to cutout, through a small number of user interaction as less as possible.In most of the photo editing software, users are required to find the contour by tracking the entire contour of the object. It is very tedious, and needs lots of user-interface, but the result is not always satisfied. There have been some better algorithms such as Snake model and Lazy snapping. Snake model needs a coarse contour and Lazy snapping only need some scribbles to distinguish background and foreground. However, Snake model cannot handle many types of image because it needs many parameters which is hard to decide, and highly depends on initial contour. Lazy snapping might fail when the foreground and background share the similar texture.We propose a novel image cutout algorithm based on local similarity. The user interface is different from Snake and Lazy snapping. The users only mark some critical points along the boundary of the object. Then the whole contour of the object is extracted from these critical points based on the local similarity through an optimization process. First, we will introduce the algorithm based on similarity of normal vector neighbor, and then we improve on similarity of local neighbor. We also improve on the user-interface of marking points along boundary, which let users not have to accurately mark points along the boundary. That means to allow same tiny error, so it makes the user-interface easy.This algorithm needs little user-interface, and the user is just required to mark only a few points along the contour of the object. The results are better than the previous algorithms. It can deal with the image has similar texture between foreground and background which Lazy snapping does not work well. If we mark enough number of points along the boundary we can get more exact results. Finally, the algorithm is fast and can work on a variety of different images.